Anthropic splits its new flagship model in two — and only one version is for you
Claude Fable 5.1 is cheaper and generally available; its sibling Mythos 5.1, with fewer safety brakes, is locked behind vetting that most people will never see.
Anthropic has announced two new models, Claude Fable 5.1 and Claude Mythos 5.1, which the company bills as its “world’s most advanced” for coding and knowledge work. Strip away the marketing gloss, though, and the headline detail is simpler: it’s the same underlying model wearing two different sets of guardrails, and only one of them is actually available to ordinary users.
What’s actually new here
Fable 5.1 is the general-release version, and Anthropic says it costs roughly 25% less than Fable 5 for typical billed usage, rising to as much as 45% cheaper for heavily “agentic” workloads — tasks where the model runs long, multi-step processes rather than answering single prompts. The saving comes from cheaper pricing on cached inputs the model has already processed. That’s Anthropic’s own estimate, not an independently audited figure, but it’s a concrete and checkable claim once real invoices start rolling in.
Mythos 5.1, by contrast, is the same model with lighter safety restrictions, and it isn’t available to the public at all. Anthropic says it’s reserved for “trusted access programmes” aimed specifically at cybersecurity and life-sciences work, including a biology-focused access scheme built in partnership with the US government, with enrolment for scientists said to be opening “soon”. For everyone else, Mythos simply doesn’t exist as an option.
The privacy and safety changes worth checking
Anthropic is also introducing “Enterprise Frontier Safeguards” (EFS), a system that stores customer data on infrastructure the customer controls rather than Anthropic’s own servers, which the company says delivers privacy equivalent to zero data retention while still catching abuse. That’s a meaningful claim if it holds up, but it isn’t live yet — Anthropic says it’ll roll out to enterprise customers “in phases” starting later this autumn. Until then, eligible customers get a standard zero-data-retention option instead, which is a narrower and already-familiar safeguard.
On the security side, Anthropic claims its cybersecurity filters now throw up 60% fewer false positives, partly because Fable 5.1 has been tuned to help find software vulnerabilities without helping develop exploits for them. That distinction is doing a lot of work here — it’s a policy line drawn by the model’s training and safeguards, not a hard technical wall, so it’s worth treating as a stated intention rather than a guarantee.
So who does this actually affect
If you use Claude Code or similar tools through the standard consumer or developer channels, Fable 5.1 is what you’ll get, and the practical upshot is lower token costs, particularly for long-running automated tasks. Mythos 5.1 is not something most readers will ever touch; it’s aimed at vetted researchers and security professionals working inside formal access schemes.
On the Hacker News discussion attached to the announcement, an Anthropic employee highlighted improvements to writing style and a doubling of the model’s score on a specialised science benchmark, framing it as an early sign of AI contributing to research. That’s an interesting internal metric, but a benchmark score is not the same as a scientific breakthrough, and readers should treat “early glimpse of scientific progress” as Anthropic’s framing rather than an established fact.
The takeaway
The concrete, checkable news is a price cut and a genuinely tiered access system — cheaper AI for most people, tighter-but-relaxed AI for a gated few. The loftier claims about being the “most advanced” model or a step toward AI-driven science are Anthropic’s own assessment, worth watching rather than taking at face value until independent benchmarks and real-world use catch up.